Machine learning techniques for breast cancer diagnosis

Full Text

INTRODUCTION

In 2022, breast cancer ranked first among all detected cancers, accounting for 19% [1]. Early diagnosis is critical, contributing significantly to successful treatment and enhancing survival. In recent years, machine learning (ML) has gained increasing attention for applications in diagnostics, including breast cancer. Zou et al. compared ML techniques using patient demographics, medical records, imaging data (mammography, ultrasound, magnetic resonance imaging [MRI]), and genetic profiles as input information [2]. The results indicated that ML algorithms such as Support Vector Machines (SVMs), Random Forests (RFs), Artificial Neural Networks (ANNs), and deep learning models (DLMs) accurately diagnosed breast cancer. The study identified the potential of ML in diagnosing breast cancer. This review discusses ML techniques used to solve auxiliary tasks in current breast cancer diagnostics, such as intelligent patient status assessment for preliminary diagnostic decisions and as predictive systems.

USE OF MACHINE LEARNING TECHNIQUES TO SOLVE AUXILIARY TASKS IN MEDICAL DIAGNOSTICS OF BREAST CANCER

ML can be effectively utilized to solve auxiliary tasks during breast cancer diagnosis. Image segmentation using computer vision is one such task. ML techniques such as Convolutional Neural Networks (CNNs) can automatically segment tumor images and other mammographic abnormalities [3]. Another task is to classify the data to determine the presence or absence of breast cancer. ML techniques such as SVMs [4] or RFs [5] can be applied. In this step, approximately 120 studies that used ML models to automate routine tasks or solve auxiliary ones were analyzed. Below are some examples.

In addition, ML can be used to analyze genetic information to predict breast cancer based on genomic data [6]. We reviewed published studies that applied ML to these tasks in breast cancer diagnosis using traditional diagnostic techniques such as mammography, ultrasound, and MRI.

One of the first studies to use ML to diagnose breast cancer used a neural network [7]. Surface breast density and the coefficient of variation for the luminal form factor were the input data. The neural network correctly classified 92% of the cases in the dataset, and the authors suggested that neural networks may be helpful in differential diagnosis. This breakthrough study paved the way for advancing ML in breast cancer diagnosis. Since then, ML techniques have continued to evolve, improving the accuracy and speed of diagnosis and reducing the number of false results.

Machine learning techniques for mammography

ML algorithms are actively applied in the established modalities of breast cancer diagnosis. For example, a study used a CNN to reduce the number of false positives and classify tumors as benign or malignant [8]. ML has provided high diagnostic accuracy and improved imaging efficiency. Another study utilized a CNN to localize a tumor, even in multiple locations, and classify it as benign or malignant [3]. Fig. 1 shows an example of the use of such a model. The resulting classifier had an accuracy of approximately 85% for test examples, serving as a tool to assist radiologists in detecting breast cancer. A similar study employed a radial basis function neural network to classify and segment breast tumors automatically [9]. It demonstrated high accuracy of tumor localization using the original input image with color results. Another study utilized a technique to classify breast lesions in digital mammography with an accuracy of 95.8% [10]. Additionally, a study aimed to test trained models on a dataset that differed markedly from the one used for training [11]. Data were preprocessed by normalizing image channels values within the range of 0 to 1. The study identified a problem in classifying images with alterations in color balance, contrast, and brightness, which markedly reduced accuracy, average recall, and the ROC curve.

Fig. 1. Example of tumor location [3].

Machine learning techniques in ultrasound

Despite the high accuracy of mammography in cancer diagnosis, this modality has a serious limitation of exposure to ionizing radiation. Therefore, it is used only for women over 40 years and is subject to additional restrictions¹. Currently, ultrasound (US) is widely utilized alongside mammography. This modality is being investigated for use with ML to improve diagnostic accuracy. US offers several advantages, including ease of use, safety, noninvasiveness, and low cost. One study described the process of breast cancer diagnosis using a US device in combination with a logistic regression model to evaluate the shape, borders, and classification of blood flow during the differential diagnosis of benign and malignant tumors [12]. It resulted in a model with an increased diagnostic accuracy of 92.4%. Another study built a neural network model and applied it to segment breast cancer US scans [13], with an accuracy of 97.3%. However, the dataset was small (90 patients), which may have unpredictable impacts on the results during validation. A study applied a convolutional neural network to US measurements to detect early-stage breast cancer [14]. Fig. 2 shows the images of two patients. Each example shows a grayscale US image and the corresponding heat map. The red area represents more weight, which can be decoded by applying the color scale on the right. Image “a” indicates that the low-frequency region within the tumor is valuable for predicting the status of the perivascular lymph nodes, whereas, in image “b,” the tumor border is indicated. However, US has its limitations; the results may depend on the sonographer’s expertise and skills, and small or early-stage breast tumors may not always be detected [15].

Fig. 2. Visualization of two patient cases [14]: Image “a” shows that the low-level region within the tumor is valuable for predicting the status of perivascular lymph nodes; Image “b” indicates the tumor border.

Machine learning in magnetic resonance imaging

Intravenous contrast-enhanced magnetic resonance imaging (MRI) is currently the most accurate technique to detect breast cancer. MRI detects cancer in 96% of cases, which is relatively high compared with mammography or US [16]. ML is also being developed to improve diagnostic accuracy in this field. A study used MRI-based deep learning to calculate the probability of breast cancer, with the proposed system having a ROC AUC of 0.92 [17]. A similar study used a CNN to classify MRI images with a network accuracy of 98.33% and an error rate of 0.0167 [18]. Another study proposed a way to normalize MRI scans as images produced by different MRI scanners varied in intensity and noise distribution, making it difficult for algorithms trained on images from one scanner to generalize data from others [19]. A cyclically consistent, generative, adversarial network was used. This approach normalized images and could potentially improve the diagnosis and detection of breast cancer. However, this modality has its limitations; it is a rather expensive procedure that requires highly qualified specialists and access to scanners [20].

PRELIMINARY EVALUATION OF BREAST CANCER DIAGNOSIS USING MACHINE LEARNING

In addition to these diagnostic methods, certain approaches identify risk factors and use them for a more thorough screening [21–23]. Computerized microwave radiothermometry is one such modality [21]. This computerized instrument is designed to measure the intensity of the ultrahigh-frequency electromagnetic radiation naturally emitted by the patient’s internal tissues; the intensity is proportional to the tissue temperature. It is utilized in oncology, neurology, urology, and gynecology. Its main advantages are its safety, having no contraindications, and applicability in diagnosing and monitoring the treatment of various diseases. Fig. 3 shows a thermogram of a healthy subject. However, the analysis and interpretation of thermometric data is challenging [24].

Fig. 3. Thermogram of a healthy subject.

Thermography is another modality that is based on the measurement of thermal radiation with an infrared camera. It records and converts infrared radiation into an image, called a thermogram, which indicates heat distribution on the body’s surface [25]. Thermal scintigraphy is another modality that uses a radioactive compound [26]. An infrared camera detects its accumulation in cancer cells to locate the tumor on an image.

In addition to thermal data processing, ML models also employ breast cancer-related genetic data as input. These modalities can help diagnose early-stage breast cancer by identifying problems cost-effectively and referring a patient for traditional diagnostic testing if needed. However, they do not provide a complete picture and do not allow for accurate diagnosis because an increase in temperature or mutation of specific genes is not always associated with breast cancer. As data interpretation is difficult, ML is used to identify risk factors and produce information, understandable to a healthcare professional and provides additional arguments for further examination. In this step, approximately 90 studies were analyzed that used ML models for breast cancer screening. Below are some examples.

Microwave radiothermometry

A study used a weight-insensitive neural network to detect breast cancer using microwave radiometry (MRT) [27]. The model had an accuracy of 92%. Another study utilized a genetic algorithm to localize malignant breast tumors based on MRT data [28]. Fig. 4 indicates the 18 breast sites measured for temperature. The algorithm results had an accuracy of 55%–65% with the test sets. A study proposed an intelligent guidance system to quickly navigate the temperature field changes during diagnosis [29]. These studies described the potential of MRT while using an ML-based diagnostic tool to detect cancer risk. However, this modality has some limitations. It is necessary to prepare a list of features to be used in advance because not all data obtained may apply to cancer diagnosis [23]. In addition, MRT detects only thermal changes, which depend primarily on the tumor growth rate and is ineffective at low rates.

Fig. 4. Locations of temperature measurement points.

Thermography

A study used thermography (TG) to detect breast cancer based on ML with an ROC AUC of 0.89 [30]. Another study proposed a system based on TG and an SVM to detect breast cancer [4], achieving an accuracy of 96.57%. The study also demonstrated a classification accuracy of 92.70% compared to 82.05% for mammography while differentiating between benign and malignant tumors. A similar study used an InceptionV3 CNN to analyze the breast cancer thermal images [31]. The resulting model classified images as abnormal or healthy with a reliability of 0.78 or 0.94, respectively. A study utilized infrared thermometry within the SPSS environment with ML techniques [32]. The sensitivity and specificity of the TG applied in this study were 89.9% and 76.4% in women ≤55 years and 90.3% and 78.9% in women >55 years, respectively. Despite its high efficiency, TG cannot visualize anatomical structures and only detects infrared thermal radiation emitted from the skin surface, which is not always a contributing factor in breast cancer development.

Genetic testing

The study used ML algorithms such as the generalized linear model, RF, gradient boosting, and DLM to identify complex patterns in germline DNA length that correlated with breast cancer [5]. Changes in germline DNA length were measured at the chromosomal level. These measurements represented the sum of multiple chromosomal insertions, deletions, and copy number changes. The result was a gradient-boosting software product with an ROC AUC of 0.83 for binary classification. Another study applied lifestyle and reproductive factors along with genetic profile information in a logistic regression model to identify breast cancer risk [2]. The model had a ROC AUC of 0.73 and was considered a starting point for breast cancer screening. A study proposed a breast cancer classification system based on gene expression [33]. Several ML techniques, such as a recurrent neural network, k-nearest neighbors, a probabilistic neural network, a CNN, and genetic algorithms were utilized. The best classification results were selected by combining each technique using a genetic algorithm and comparing them for accuracy. The accuracy of the proposed technique was 97%. A study proposed a model combining gradient-boosting algorithms and explicable artificial intelligence to predict breast cancer metastases based on genomic data [6]. The result was a model with 96% accuracy and 99.3% ROC AUC. This diagnostic modality has some advantages over others, such as high accuracy (detecting specific genetic mutations probably associated with breast cancer) and early detection (identifying genetic mutations that may predispose to breast cancer, even before the first symptoms appear). However, some limitations included high costs and limited data. Not all mutations in breast cancer genes are currently known, restricting the ability of genetic analysis to fully and accurately identify all risks and signs of disease progression. However, they can provide a basis for designing more accurate diagnostic modalities.

Linguistic description of the patient’s condition

In addition to using ML to analyze clinical findings, approaches also use medical history, ultrasound results, and other linguistic parameters as input information. Magna et al. applied deep learning techniques to a recommendation system for breast cancer detection utilizing patient records [34]. Input information included clinical anatomy, disease type, physician consultation data, procedures, and medications. In total, 20 experiments with 5-fold cross-validation yielded an average accuracy and recall of 98% for the classification of “cancer” versus “not cancer” and 98.6% for “breast cancer” versus “other cancer.” Another study employed family history and patient questionnaire replies as input information [35]. A total of 603 answers were collected, including data from 309 patients with breast cancer and 294 healthy subjects. Another study described data collection and analysis methods in detail [36]; three classifier algorithms were evaluated as breast cancer prediction models:

• SVM;
• RF;
• Multilayer Perceptron (MLP).

The evaluated RF, SVM, and MLP models reliably classified breast cancer and healthy cases with an average sensitivity of >97.2%, specificity of >96.4%, and accuracy of >97.1%.

Afrash et al. analyzed data from 3,168 healthy subjects and the medical records of 1,742 patients in an Iranian hospital [37]. The variables selected to predict breast cancer included age, dairy consumption, family history, breast biopsy, chest X-ray, hormone therapy, alcohol consumption, being overweight, having children, and education status. The experiment showed that Decision Tree performed better than the other ML models, with accuracy, specificity, and sensitivity of 99.3%, 99.5%, and 98.26%, respectively.

CONCLUSIONS

This study analyzed more than 200 open-source scientific publications. Papers from Russia, the USA, China, Iraq, etc. were included. Fig. 5 shows a diagram of the most commonly used ML models for breast cancer diagnosis. ML techniques can solve two main tasks: auxiliary tasks during the medical diagnosis of breast cancer and preassessment of breast cancer diagnosis. For the first task, we can mention studies aimed at classifying tissue samples as malignant or benign. Additionally, ML can be used to identify risk factors for breast cancer. Another example is the analysis of molecular data, such as genomic data. Table 1 summarizes the ML techniques employed to solve such tasks. The application of ML techniques in breast cancer diagnosis offers great opportunities to improve diagnostic accuracy and efficiency and solve ancillary tasks. For example, CNNs can recognize complex hierarchical features in images. Thus, the neural network can automatically learn features such as thermal field asymmetry, heterogeneity, microcalcifications, etc., which are key to diagnosing breast cancer, allowing the detection of pathological changes invisible to the human eye. The advantage of a probabilistic neural network in mammography is its ability to handle uncertainty and consider the probabilities of different classes. This ability provides a more appropriate and reliable diagnosis, especially in ambiguous and complex cases. One of the key advantages of deep learning is its ability to extract complex hierarchical features automatically from large amounts of data. This capability may be particularly helpful for predicting breast cancer progression, as the medical data utilized for diagnosis have many features that can be difficult for humans to perceive. A model that may aid in breast cancer diagnosis was created utilizing logistic regression to classify blood flow based on US data. Its main advantages were its robustness to outliers and data noise, meaning that it can handle US data that may contain noise or imperfections without significantly affecting the classification accuracy. An SVM was used to classify breast cancers based on thermal data, effectively when training data was limited. This ability is specifically useful when a limited number of breast cancer cases are available for model training. Identifying the key genetic markers or features that may be associated with a predisposition to breast cancer is one of the most crucial aspects of breast cancer diagnosis. Genetic algorithms can effectively process large amounts of data, especially because much clinical and genetic data are available for breast cancer classification. However, the limitations of these techniques, such as the need for a large amount of training data as well as verification and validation of the results obtained, should be considered. An analysis of the available sources concludes that medical diagnostics actively use computer technologies, especially ML techniques such as neural networks, regression, and RFs. The studies discussed in this review clearly indicate that ML can improve the screening of patients; accuracy and adequacy of diagnosis; automatically classify tumors, and prepare data for further processing.

Fig. 5. Rates of machine learning models usage (%).

Table 1. Machine learning and its possible uses in breast cancer diagnosis

ADDITIONAL INFORMATION

Funding source. This article was not supported by any external sources of funding.

Competing interests. The authors declare that they have no competing interests.

Authors’ contribution. All authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work. I.V. Germashev — contribution to the development of the research concept, methodology, and overall project management; K.S. Dyomin — data collection and analysis, writing the main sections of the article.

¹ Order of the Ministry of Health of the Russian Federation No. 404н (404n) dated April 27, 2021 (as amended on February 1, 2022) “On Approval of the Procedure of Preventive Medical Examination and Screening of Certain Groups of the Adult Population.” Available at: https://rcmp-nso.ru/profila/m_mater/docs1/order_rf404n.pdf?ysclid=m0q99isd18984245105 Accessed on: January 9, 2024.

About the authors

Kirill S. Dyomin

Author for correspondence.
Email: diominkirill@yandex.ru
ORCID iD: 0009-0002-4571-3437
Russian Federation, Volgograd

Ilya V. Germashev

Email: germashev@volsu.ru
ORCID iD: 0000-0001-5507-8508
SPIN-code: 2489-2628

Dr. Sci. (Engineering)

References

Supplementary files

Supplementary Files

Action

1. JATS XML

Download

2. Fig. 1. Example of tumor location [3].

Download (118KB)

Indexing metadata

3. Fig. 2. Visualization of two patient cases [14]: Image “a” shows that the low-level region within the tumor is valuable for predicting the status of perivascular lymph nodes; Image “b” indicates the tumor border.

Download (215KB)

Indexing metadata

4. Fig. 3. Thermogram of a healthy subject.

Download (283KB)

Indexing metadata

5. Fig. 4. Locations of temperature measurement points.

Download (69KB)

Indexing metadata

6. Fig. 5. Rates of machine learning models usage (%).

Download (107KB)

Indexing metadata